OpenAI Codex gains device agency

OpenAI expanded Codex’s control over Macs on May 21, adding features that let the coding agent keep working after a device locks, pull context directly from the screen and continue longer-running tasks. The changes were disclosed in OpenAI’s Codex changelog and ChatGPT release notes, and were picked up by Macworld and Cult of Mac. The update adds “Appshots” in the macOS app, makes Goal mode generally available across the app, IDE extension and CLI, and enables “locked computer use” for eligible Mac users under existing regional constraints. OpenAI’s developer docs say remote connections also let users steer Codex from another device, including the ChatGPT mobile app or an SSH-connected machine. ### How is Codex getting context from the screen? (developers.openai.com) OpenAI said Appshots let a user press both Command keys to send the frontmost app window to a Codex thread, including a screenshot and “available text.” That means the model can work from what is visible in another app without the user copying code, pasting logs or writing a long prompt to describe the screen. Cult of Mac described the feature as giving Codex for Mac “eyes,” while OpenAI’s own documentation frames it as a way to attach app context directly from macOS. (developers.openai.com) The feature is limited to the frontmost window rather than a full desktop recording, according to the changelog language. ### What does “locked computer use” actually allow? OpenAI’s May 21 release notes said “locked computer use” lets eligible Mac Computer Use users keep Codex working remotely and securely after the Mac locks. (developers.openai.com) OpenAI’s Computer Use documentation says the feature requires installation of an Apple authorization plug-in that participates in the macOS unlock flow. OpenAI’s enterprise release notes said admins and owners can turn the feature off by setting `remote_computer_use = false` in Codex cloud policies and configurations. (cultofmac.com) Macworld reported the same capability as Codex being able to control a Mac even when it is locked. ### Where does remote control fit in? (help.openai.com) OpenAI’s remote-connections documentation says Codex can be used from another device or machine, including a connected Mac, another Codex App device or projects on an SSH host. The Codex app page also says users can start, steer, approve and review work from the ChatGPT mobile app on a connected host. That matters because the new release bundles remote access with longer-running execution. (help.openai.com) OpenAI said Goal mode is no longer experimental and can drive toward a specific objective “for hours or even days.” ### What permissions and controls still sit around the agent? OpenAI’s Computer Use documentation says Codex needs Screen Recording permission to see a target app and Accessibility permission to click, type and navigate. (developers.openai.com) The same documentation says those macOS permissions are separate from Codex app approvals, while file reads, file edits and shell commands continue to follow sandbox and approval settings for each thread. (developers.openai.com) OpenAI also says users should prefer a dedicated plugin or MCP server when an app exposes one, reserving Computer Use for cases where Codex needs to inspect or operate software visually. That language suggests OpenAI is distinguishing between structured tool access and direct interface control. ### What changed in practical terms for engineering teams? OpenAI’s April product post said the updated Codex app adds computer use, in-app browsing, memory and plugins to accelerate developer workflows. (developers.openai.com) The May 21 release extends that by letting Codex ingest screen context faster, continue after lock and pursue longer goals, moving more work from chat-style assistance into machine operation. That last point is an inference from the feature set, based on OpenAI’s product and documentation language. For engineering and IT teams, the immediate operational facts are narrower: Codex can now see more context from macOS windows, act through Accessibility and Screen Recording permissions, and keep running remotely after lock if the feature is allowed. The documented control points are per-thread approvals, sandbox settings and an admin policy switch for locked remote use. (help.openai.com) (openai.com)